Fix monkeypatch undo when deleting missing attrs/items with raising=False

[kartikdp]

Summary

Fixes a state-restoration bug in MonkeyPatch when deleting a missing attribute/item with raising=False.

Previously:

• delattr(..., raising=False) and delitem(..., raising=False) did not record the missing state.
• If the test then added that attribute/item, undo() would not reliably restore the original missing state.

This change records NOTSET for those delete calls, and makes undo() tolerant when the target attribute is already absent.

Closes #14094.

Why this matters

Tests that use non-raising deletes expect monkeypatch teardown to fully restore pre-test state. Without this fix, state could leak after tests.

How tested

• uv run -m pytest testing/test_monkeypatch.py -k "delattr_non_existing_with_raising_false or delitem_non_existing_with_raising_false or test_delattr or test_delitem"
• uv run -m pytest testing/test_monkeypatch.py
• uv run --with pre-commit pre-commit run --files src/_pytest/monkeypatch.py testing/test_monkeypatch.py changelog/14094.bugfix.rst

[nicoddemus]

This test already passes on main, so it is not reflecting/reproducing the issue.

Did you intend to use setattr directly to set x (instead of monkeypatch.setattr).

[nicoddemus]

Same comment as the other test.

[nicoddemus]

Not sure if this is a bug-fix, seems like an improvement: now delattr will always delete the attribute at the end of the test.

This is a behavior change which I don't think can cause problems later, as it depends on the user explicitly calling monkeydelattr.delattr(..., raising=False).

[nicoddemus]

@The-Compiler you commented on the issue; are you OK with this?

[The-Compiler]

I am indeed -1 on this. Here's a trivial example:

import types


def test_attribute(monkeypatch):
    obj = types.SimpleNamespace()
    with monkeypatch.context() as mp:
        mp.delattr(obj, "attr", raising=False)
        obj.attr = 42
    assert hasattr(obj, "attr")

It seems natural to me to assume that the object still has the attribute after monkeypatch is done. After all, I never set the attribute through monkeypatch, so I really wouldn't expect monkeypatch to mess with it (as, with raising=False, it was a no-op). I would be fine if there was an opt-in way with an undo=True or something, that also opens up the opportunity to document this quirk in a nice way. Right now, with this PR, this seems like a footgun to me.